Meta's Iris Chip Enters Production in September. Broadcom Is Quietly Winning the Custom Silicon Race.

Reuters got the memo on Thursday. Meta is putting the Iris chip, the first production silicon out of its MTIA program, into manufacturing in September. Broadcom did the design work. TSMC has the fab allocation. Iris cleared its bug-testing phase in about six weeks with no significant issues, which is roughly the fastest anyone has taken a custom AI accelerator from tape-out to production this cycle. The memo also spelled out the compute-capacity target and the CapEx line: seven gigawatts online in 2026, fourteen gigawatts by 2027, up to $145 billion of AI infrastructure spend for the year.

The headline read is that Meta joined the custom-silicon club. That is the surface story. The one worth pulling on is that Broadcom just added a second $100B-plus program to the same design bench that already carries TPU v7 for Google, which is the silicon underneath the $200 billion Anthropic commitment we wrote up in May. The same company is quietly on both sides of the largest ASIC buildouts in the industry.

The Iris Memo In Numbers

Two of those numbers belong in a case study. Six weeks of bug testing is not normal for a new-node accelerator. Google spent close to a year on TPU v5 validation before green-lighting fab starts. Amazon needed most of a year on Trainium 2. Six weeks tells you Iris is either (a) an architecturally conservative refresh of the MTIA v1 die with a Broadcom-tuned interconnect, or (b) validated primarily against Meta's own known workloads, which have narrower coverage than a general-purpose accelerator. Both readings are more likely than the third possibility, that Meta simply moved faster than the rest of the industry on a fresh architecture. The memo is not describing a moonshot. It is describing a chip Meta is confident enough in to commit fab and packaging capacity against.

The other number that matters is the CapEx raise. Meta had been guiding investors to roughly $118 billion of 2026 infrastructure spend at the last earnings call. The memo puts the top of the range at $145 billion. That is a $27 billion swing inside one calendar quarter, which is bigger than the annual R&D budget of every AI lab except OpenAI and Anthropic. The compute-buildout arms race we sketched out in the buildout piece just got a fresh delta from the largest ad platform on the internet.

The Broadcom Story Nobody Is Telling

Every ASIC needs a design partner. That is where the chip actually gets turned from a workload spec into physical silicon. The current tally at the top of the market:

Broadcom is now on three of the five most important custom AI programs in the industry. Google TPU has been the flagship for years. OpenAI Jalapeno, which we covered when the deal closed and wrote up as the closed silicon loop, added the second flagship in Q1. Meta MTIA/Iris makes it three. The company's AI revenue guide for fiscal 2026 was $17 billion at the last update; the run rate implied by these three programs alone, on multi-year contracts, is closer to $60 billion by 2028.

Broadcom trades at roughly 22 times forward revenue right now. If the market caught up to the ASIC concentration story, it would be trading closer to Nvidia's multiple. It is not. The market still prices Broadcom as a networking silicon business with an AI attachment. The reality is closer to inverted. Broadcom is the ASIC business the market has not repriced yet.

What This Does to Nvidia

Less than the ASIC-shipment-growth headlines suggest, but more than zero, and the composition of the delta matters. Custom AI chip shipments are on track to grow about 45 percent in 2026 against roughly 16 percent for merchant GPU shipments, per Futurum Research's July update. That gap is the sharpest single data point on where marginal silicon dollars are going.

Nvidia is not losing revenue on any of this. Its 2026 data-center number is still tracking to a record. What it is losing is future customer concentration. Until this year the bull case on Nvidia assumed Meta, Microsoft, Google, and Amazon would each keep buying roughly a fifth of every new node's output for the foreseeable buildout. That assumption is now empirically wrong for three of the four. Google is buying half as many Nvidia GPUs per gigawatt as it was in 2024, because TPU covers the rest. Microsoft is publicly guiding to Maia 200 handling the Copilot inference base, which we covered in the Maia 200 piece. Meta with Iris is the last of the top-four hyperscalers to lock in the split.

The single largest remaining Nvidia-dependent frontier buyer is now OpenAI, which still runs Vera Rubin on Azure and CoreWeave as its dominant training substrate and does not take Jalapeno silicon at meaningful volume until 2027. Anthropic is the second largest, and Anthropic's $200B commitment is now TPU-anchored. That is the shape of Nvidia's 2027 customer file: two labs on top, hyperscalers filling in around the edges, and the growth math meaningfully harder than it was six months ago.

The 14 GW Question

Doubling data-center compute from 7 gigawatts to 14 gigawatts inside twelve months is not primarily a silicon problem. It is a power problem, a substation problem, and a permitting problem. Meta already has the majority of its 2027 gigawatts under signed power purchase agreements, but the transformer and switchgear lead times sit at eighteen to twenty-four months for anything above 300 MVA, and the FERC interconnection queue we wrote up in April has not shortened. Meta's answer, reportedly, is a mix of Louisiana natural gas co-siting, some behind-the-meter solar, and at least one nuclear PPA under negotiation. None of that is in the memo. All of it is implied by the number.

The 2027 delivery window keeps hardening into a real cliff. Anthropic's TPU capacity arrives in 2027. OpenAI's 10 GW Vera Rubin commitment on Nvidia arrives starting in 2027. Meta's second 7 GW arrives in 2027. Add Microsoft's Azure expansion and the SAP Prior Labs European buildout and you get roughly 40 gigawatts of new AI compute landing on the grid inside one year. The industry is now betting the same twelve-month window as if it were the launch pad for the whole next decade of scaling.

Our Take

The clean read on the Iris memo is that Meta closed the fourth hyperscaler ASIC loop. Google closed it with TPU. Amazon closed it with Trainium. Microsoft closed it with Maia 200. Meta's Iris makes four for four. That is the concrete end of a strategic transition every big infrastructure buyer has been running since 2023, and it happened without a keynote. It happened in a Reuters memo leak.

The more interesting read is that Broadcom is the invisible platform underneath it. If you are underwriting AI infrastructure exposure and you own Nvidia against the buildout, you are half-hedged at best. The other half of the trade is the design partner sitting on the call sheet for three of the five biggest ASIC programs in the industry. Broadcom does not make the chip in a way retail investors instinctively price for. It makes the chip possible. In 2027 that is going to be reflected in the multiple, and by the time it is, the story will be finished.

Practical implication for builders. Meta is not selling MTIA capacity externally. Iris matters to you only as a floor on internal Facebook and Instagram AI cost, which sets the shape of the free-tier consumer AI product Meta ships next. If Meta's inference cost per token on its own chip lands anywhere close to the TPU curve, expect Muse Spark 1.1 and its successors to be priced aggressively into the same commodity tier that GPT-5.6 Luna and Grok 4.5 opened up last week in the coding scoreboard. The bottom of the market keeps getting cheaper. The design partner on the silicon that makes it cheap is Broadcom. That is the actual story of the Iris memo, and it does not fit inside a headline.

We are tracking hyperscaler chip cadence on our Meta provider page and the ASIC-versus-GPU capital rotation on the inference-money piece. Next data point to watch: whether Meta lets an outside customer touch Iris at all, and whether Broadcom guides Q3 AI revenue up on the design-win pipeline. Either would confirm the story the market has not yet priced.